Multilingual Mention-Weighted Network of Wikipedia Biographies

Each node is a person, identified by their Wikidata entity ID, selected because the Cross-Verified Database records a Wikipedia biography in all five covered language editions (a small number of pages did not resolve at retrieval — see Files). Each directed edge from node u to node v in language layer l records how prominently v appears in u's biography in that edition, measured as the number of times v's name is detected in the body text. The five language editions (English, French, German, Italian, Spanish) constitute five network layers. Because the same Wikidata ID resolves to the same person across editions, nodes that are present in more than one layer are aligned across them without any name matching: the Wikidata ID is the shared key. The node set is a subset of the Cross-Verified Database of Notable People [1], which covers individuals born between 3500 BC and 2018 AD, filtered to those with biography pages in all five editions (see Files section). Edges are based on Wikipedia hyperlinks present in biography body text as of January 2, 2026, weighted by NLP-based mention counts. This dataset was built as input to a multilayer network backboning study investigating inclusion bias in Wikipedia biographies across gender, historical period, and geographic origin. Files entities_filtered_by_languages.csv — 94,403 rows, one per node. The Cross-Verified Database [1] contains roughly 2.29 million notable people. This file is a subset of those rows, keeping only the 94,403 entities that the Cross-Verified Database records as having a Wikipedia biography page in all five covered language editions. That filter is what makes a near-aligned multilayer network possible: by selection, almost every node exists in every layer. The alignment is not perfect, because the filter is based on the editions declared in the source database, whereas the layers are built from pages actually retrieved from the Wikipedia API on January 2, 2026. Of the 472,015 possible (entity, edition) page slots, 377 did not resolve at retrieval time (pages moved, deleted, or with stale sitelinks), so 327 of the 94,403 nodes (0.35%) are missing a page in at least one edition; 94,076 have a page in all five. The metadata columns are taken directly from the source database without modification. Column Type Description wikidata_id string Wikidata entity ID, Q-prefixed (e.g. Q42) birth float Birth year in the proleptic Gregorian calendar (negative = BCE); missing for 2,592 entities bigperiod_birth string Coarse historical period derived from birth year (see values below) un_subregion string UN M.49 geographic subregion of birth place; missing for 328 entities gender string Male (76,948) / Female (17,403) / Other (52) bigperiod_birth takes one of six ordered values: 1.Ancient History Before 500AD, 2.Post-Classical History 501-1500AD, 3.Early Modern Period 1501-1750AD, 4.Mid Modern Period 1751-1900AD, 5.Contemporary period 1901-2020AD, Missing. SpotlightWeightSource_0102_0505_fullmatch_corrected.csv — 12,444,180 rows, one per candidate edge per language layer. Column Type Description language_code string Wikipedia language edition: en, fr, de, it, es source_wikidata_id string Wikidata ID of the person whose biography contains the mention target_wikidata_id string Wikidata ID of the person being mentioned weight float Final edge weight after correction (see below) The candidate edge set was built by querying the Wikipedia API (snapshot of January 2, 2026) for all hyperlinks present in each biography page, regardless of where in the page structure the hyperlink appears. This means the list includes links from the running body text but also from infoboxes, footnotes, navigation boxes, "see also" sections, and other page elements. As a result, many of these links do not correspond to an actual mention in the narrative text and the NLP step will find nothing to count: 77.2% of rows have weight = 0.0. The remaining 2,836,525 rows have weight > 0, meaning at least one body-text mention was detected: Layer Nodes Edges (weight > 0) English (en) 83,823 776,770 French (fr) 79,176 620,310 Italian (it) 75,591 542,720 Spanish (es) 77,569 483,211 German (de) 81,377 413,514 All layers combined 92,238 2,836,525 Node counts per layer are lower than the total 94,403 for two reasons: most layer-absent nodes simply have no weight > 0 edges in that edition, and a small number (327 nodes overall) have no retrieved page in that edition at all (see the entities file note above). The table above includes self-loops (6,113 across all layers, of which 4,575 have weight > 0). The backbone-input graph used for analysis additionally removes self-loops, yielding 2,831,950 edges and 92,197 nodes across all layers. SpotlightWeightSource_0102_0505_fullmatch.csv — same 12,444,180 rows, with an additional fullmatch_count (int) column: the number of exact full-name substring matches of the target's Wikipedia title in the source biography text. Included for transparency; the corrected file's weights are derived from it for flagged edges. graph.db — DuckDB analytical database; snapshot of the Wikipedia API retrieval of January 2, 2026. This is the pipeline's primary cache. It stores the raw API payloads (page titles, article body text, hyperlinks, pageviews) together with the derived edge weights, all keyed by Wikidata entity ID. Six tables: Table Rows Description wikidata_entity 94,403 One row per node in V; primary key is the Wikidata integer ID (Q prefix stripped for compact storage). wiki_page 471,638 Maps each (entity, language) pair to its Wikipedia page title; one row per combination. extract 471,636 Plain-text article body retrieved via prop=extracts&explaintext=1; input to DBpedia Spotlight. Two pages returned an empty extract and have no row. page_link 55,373,085 Raw directed hyperlinks across all five language layers, including links to non-notable targets. The boolean notable column flags edges whose both endpoints are in V. weights 12,444,180 Spotlight mention-count weights keyed by (link_id, method); method = 'SpotlightWeightSource' throughout this dataset. These are the pre-correction weights; the corrected values are in the CSV files above. pageview 471,638 Daily-average Wikipedia pageviews per (page, language) over 2023-01-01 – 2025-12-31 (1,096 days), sourced from the Wikimedia REST API. Wikidata IDs are stored as integers internally (Q prefix stripped) to reduce index size; all joins across tables use this integer key. How weights are produced NLP mention count. DBpedia Spotlight [2] (confidence threshold 0.8, one local instance per language) annotates each biography's body text and counts how many times the target entity is recognised. This is the initial weight. Full-name match count. Independently of Spotlight, the target's Wikipedia title (with parenthesised disambiguators stripped) is counted as a literal substring in the source text. This is stored as fullmatch_count and is less sensitive but less susceptible to name confusion. Spurious edge correction. Spotlight sometimes inflates edge weights when two people share a surname: a mention of "Darwin" in Charles Darwin's biography may be attributed to George Darwin rather than to Charles himself. These edges are flagged when both conditions hold: (1) the source and target page titles share at least one non-stopword token of length > 3, and (2) the Spotlight weight is a statistical outlier for that source node (z-score > 1.5 relative to all outgoing weights from that node). Flagged edges are corrected in one of two ways: If the target title is an ordered substring of the source title (e.g. "Napoleon" inside "Napoleon III"), weight is set to 1.0 — even the full-name match is unreliable here because the shorter name will always appear inside the longer one, but the hyperlink confirms the edge exists. Otherwise, weight is replaced by fullmatch_count, which requires the full name to match and is a more precise signal. Non-flagged edges retain their original Spotlight weight. Limitations NLP noise. The spurious edge correction catches cases where the inflation is statistically obvious, but it is not exhaustive. Edges with moderate spurious inflation (z-score < 1.5) are not corrected, nor are cases where confusion arises from reasons other than shared name tokens. The Spotlight confidence threshold (0.8) is a precision-recall trade-off; raising it reduces spurious detections at the cost of missing genuine mentions. The full pipeline can be rerun with any threshold. Language and node coverage. The node set requires each entity to have a biography page in all five covered editions simultaneously, as recorded in the source database. The Cross-Verified Database contains roughly 2.29 million entries; only 94,403 (about 4.1%) pass this filter. The five editions covered — English, French, German, Italian, Spanish — are the largest Western European language Wikipedias by speaker population. This excludes Portuguese, Russian, Arabic, Chinese, Japanese, and others. Individuals notable primarily in non-European contexts are systematically underrepresented: only 17.7% of nodes are classified as Non-Western by UN subregion (Western = {Western Europe, Northern America, Eastern Europe, Southern Europe, Northern Europe, Oceania Western World}; the 328 nodes with missing subregion are excluded from both groups). Adding further language editions is possible via the pipeline configuration; Wikidata IDs handle node alignment automatically, but each added language shrinks the joint node set since the filter requires presence in all selected editions.